基于RSOM的两阶段移动机器人闭环检测算法

Two-stage loop-closure detection algorithm for mobile robots based on RSOM tree

针对移动机器人在自主定位时出现的感知混淆和定位误差问题,提出一种基于递归自组织特征映射RSOM聚类树的移动机器人闭环检测算法。首先采用属性图模型对图像对进行相似性度量,通过对连续采集的图像序列进行分组和增量学习对不同场景进行路标建设。然后将路标中的向量投影至RSOM的各叶节点中,同时对各路标进行权值更新。最后,新算法通过两阶段检测对闭环进行判定。在第一阶段,算法利用RSOM树的快速检索能力对采集图像进行最近邻路标检索并判断该路标是否为待检测路标;在第二阶段,算法将待检测路标内所包含的所有属性图依次与采集图像进行相似性度量,最后结合阈值加权结果进行闭环检测判定。实验结果表明,该算法受环境中动态目标的影响较小,在能取得较高的召回率和准确度的同时,定位精度大幅提升。

In order to solve the perceptual aliasing and position error problems in loop-closure detection of mobile robots,a novel loopclosure detection algorithm based on RSOM tree is present. The attributed graph model is introduced to measure the similarity between a pair of images. By grouping image sequences and incremental learning,several landmarks are generated. Each landmark＇s weight is updated when feature vectors of a landmark are input into the RSOM tree. At last,the loop-closure judgment is fabricated with the by two-stage detection approach. In the first stage,the nearest landmark of the input graph is obtained,and the algorithm judge whether it fulfills the requirement of the candidate landmark. In the second stage,the similarities between each graph in the candidate landmark and the input graph are calculated. Finally,combined with the threshold weighting result,the loop-closure judgment is given. Experimental results show the new algorithm achieve good recall and precision performance with an improved positioning accuracy,since dynamic objects put a little influence on it.